This is not thread safe of course, it is not intended to be.
This was implemented as at one stage ARDOUR::init was being called multiple
times in the testsuite which was causing a subtle bug that took some
time to track down.
if sub-frames are used, frames are rounded up for the later half of
a sub-frame. e.g. @48KSPS, 25fps, 100 subframes per frame
audio-samples >=1910 already yield 00:00:00:01
bug:
- add midi track
- add additional midi port
- remove additional midi port
- press play
-> crash
because..
ARDOUR::MidiTrack::roll's route_buffer still has 2 midi-channels. This is
because processor_max_streams was not updated (no plugins added/removed)
fill_buffers_with_input() eventually asserts when trying to get the 2nd
(nonexistant) midi port via ARDOUR::PortSet::nth_midi_port() in
ARDOUR::PortSet::port().
fixes loading a session with tracks that have more input-ports
than output-ports.
the problem:
input -ports are added first, when added, IO::ensure_port() emits
IOChange::ConfigurationChanged which in turn triggers the route to
allocate corresponding output ports.
Due to this callback the output-ports were created before the actual
output-port-configuration and plugin-configuration from the
session-state was parsed and set.
do not allow port-removal if the port would be re-added immediately
after that again because the main-delivery actually needs it.
As a side effect this prevents this crash:
* create a stereo-track, then remove one output
-> unhandled exception "AudioEngine::PortRegistrationFailure&"
The problem:
- the port is removed from the RCU ports list,
but Port::drop() (which calls jack_port_unregister) is only called
from the Port's destructor at some later time.
(because a reference to the port still exists elsewhere)
- the jack-port is not yet removed.
- meanwhile Delivery::configure_io comes along and notices that
there are more audio-buffers than ports and tries to re-register the port.
- but the port still exists in jack, so it fails and throws an exception
...which is not handled.
* forward midi-data around plugins that have no MIDI-out
* allow to insert plugins with no MIDI-input at a point with one MIDI-channel
This works because excess ports (both plugin and route) remain
unconnected and use scratch-buffers.
Tested with LV2, LXVST and LADSPA.
(AU plugins with variable in/out retain the old behavior, no bypass)
fixes http://tracker.ardour.org/view.php?id=5630
Plugins rewrite the buffer data in-place and some plugins
can produce output even when fed with silence.
Hence, during a PluginInsert::silence() run a plugin can
inject data into the "silent" buffers which causes side-effects.
Kudos to Chris 'oofus' Goddard for finding this issue.
"scratch buffers are by definition scratch and their contents are undefined at all times"
"silent buffers are by definition all-zero and should not be used for real data"
But track & route were using those for actual data; plugins (which may run
in the same thread and may get the same buffers) use them for scratch thereby
overwriting real data.
In particular get_silent_buffers() (used by LadspaPlugin::connect_and_run)
clears the buffer which can holds real data:
e.g. via Route::passthru_silence() -> plugin1 -> plugin2 (clears output of plugin1)
- replaces fix in 066df0d218
- Check if _LIBCPP_VECTOR is defined for vector_delete.
This is defined in libc++'s headers which Apple is using
instead of libstdc++
